Atomoxetine’s Brain Effects Recreated by a Single Molecular Dial in Whole-Brain Model

Scientists have built a whole-brain model that draws on the cortical expression of fourteen adrenergic and dopaminergic receptor genes to reproduce, in silico, how the attention-deficit drug atomoxetine reshapes human brain dynamics. The study, published in Frontiers in Computational Neuroscience, reports that a single parameter tied to receptor binding affinity was enough to transform the model’s placebo configuration into its atomoxetine configuration, offering a computational account of how a drug that acts on one molecular target can produce distributed effects across the cortex.

Atomoxetine is a norepinephrine reuptake inhibitor that globally elevates catecholaminergic signaling. Rather than acting on a single receptor subtype, it engages a broad family of adrenergic and dopaminergic receptors with distinct binding affinities. The researchers, led by Martí Monge-Asensio of the Centre for Brain and Cognition at Universitat Pompeu Fabra in Barcelona, set out to move beyond single-receptor modeling strategies, which they argue cannot explain the collective response of multiple receptor systems to such a pharmacological perturbation.

The empirical foundation of the work came from pharmacological resting-state functional MRI data collected from 24 healthy participants, 19 of them female and 5 male, in a randomized, double-blind, placebo-controlled crossover design. Each participant completed two 8-minute open-eye scanning sessions, receiving either 40 milligrams of atomoxetine or a placebo of 125 milligrams of lactose monohydrate 90 minutes before scanning. Cortical time series were extracted using a modified Desikan-Killiany atlas with 31 regions per hemisphere, and the signals were band-pass filtered between 0.008 and 0.08 hertz.

To characterize the cortex’s catecholaminergic architecture, the team defined each cortical region by a 15-dimensional vector comprising expression values for 14 genes encoding metabotropic adrenergic and dopaminergic receptor subtypes, plus a 15th molecular gradient representing the regional excitatory-to-inhibitory receptor ratio. Pairwise receptor similarity between regions was computed, and a Leiden community detection algorithm applied to this similarity matrix identified three distinct catecholaminergic networks: frontal, posterior and temporal. These networks, the authors write, provide the neurochemical scaffold for interpreting atomoxetine-induced brain dynamics.

Empirically, atomoxetine significantly stabilized spontaneous whole-brain functional dynamics, as indicated by increased temporal correlations of functional connectivity states measured with a sliding-window approach and quantified using the Kolmogorov–Smirnov distance between condition distributions. When the researchers quantified network-specific contributions, they found that this drug-induced stabilization was localized within the posterior catecholaminergic network and in posterior-temporal interactions, while frontal and temporal networks showed no isolated within-network stabilization. The effect, in other words, was non-uniform and rooted primarily in posterior cortical circuits.

To link this pattern to receptor topology, the team constructed a multi-receptor whole-brain model in which regional neuronal gain reflects the joint modulation of excitatory and inhibitory populations by catecholaminergic receptor expression. Each region’s gain was scaled by transcriptomic estimates of receptor expression weighted by receptor-specific parameters bounded between minus one and one. Fifty such models were generated per pharmacological condition as computational replicates. Critically, these transcriptomic-based models consistently outperformed surrogate models built from spatial-autocorrelation control maps that preserved the empirical spatial variograms, and models optimized for atomoxetine better captured atomoxetine functional connectivity dynamics than placebo dynamics, which the authors describe as a necessary condition for establishing pharmacological specificity.

The model also passed a stricter test of generalization. Because the optimization targeted group averages, there was a risk that the model had merely memorized a population template. To probe this, the researchers trained linear support vector machines on functional connectivity matrices from one domain, empirical or simulated, and tested them bidirectionally across domains. Classifiers trained on empirical data successfully identified simulated drug states with a receiver operating characteristic area under the curve of 0.9 plus or minus 0.05, and classifiers trained on simulated data predicted drug conditions in individual empirical scans significantly above chance. Notably, models trained on simulated data outperformed those trained directly on empirical data when decoding drug states in empirical subjects, and spatial-autocorrelation surrogates failed this test.

The most striking result emerged when the researchers compared the fitted receptor weights between conditions. Univariate comparisons revealed no significant differences between placebo and atomoxetine models, indicating a preserved parameter profile. Yet the magnitude of the optimized receptor weights correlated strongly with empirical catecholaminergic binding affinities drawn from pharmacological databases, with a correlation coefficient of 0.83 and a 95 percent confidence interval of 0.78 to 0.89 that survived correction for multiple comparisons. The weights did not correlate with activation potency or with protein sequence similarity. This, the authors suggest, means binding affinity constrains the overall scale of receptor influence on local neuronal gain rather than its modulatory direction.

Building on this alignment, the team tested whether atomoxetine’s pharmacological identity could be reconstructed from the placebo configuration through a single global affinity-scaling transformation applied only to the 14 catecholamine-specific receptor weights, holding the atomoxetine bias and excitatory-to-inhibitory parameters at their placebo values. A systematic grid search identified an optimal scaling parameter of minus 0.15 that recapitulated the empirical static connectivity changes and dynamical shifts, including the stabilization within posterior and posterior-temporal networks. Support vector machines trained on the transformed receptor weights also discriminated placebo from atomoxetine, with negative affinity scaling systematically enhancing decoding performance. Within the model, the authors write, this affinity-governed negative reweighting serves as a plausible computational proxy for atomoxetine-induced catecholamine accumulation, in which high- and low-affinity receptors experience differential shifts that alter how the receptor ensemble collectively determines regional neuronal gain.

How this molecular redistribution is expressed across the cortex proved equally revealing. Drug-induced changes in neuronal gain did not correspond to canonical cortical hierarchy measures based on T1-weighted to T2-weighted ratio maps, nor to the empirical density of the norepinephrine transporter measured with positron emission tomography. Instead, they correlated specifically with the frontal chemoarchitectural gradient derived from the receptor similarity matrix, with a partial correlation of 0.77 and a 95 percent confidence interval of 0.68 to 0.85. A targeted virtual washout, in which frontal gain was reset to placebo baseline while drug-induced modulation was retained elsewhere, abolished whole-brain dynamical stabilization and collapsed posterior network stabilization without affecting static connectivity fits. Conversely, forward modeling that parametrically scaled the frontal gradient was sufficient to recover both static and dynamical signatures of the drug state. The authors caution that this gradient should not be read as evidence for localized frontal drug action; rather, it reflects a large-scale mechanism in which distributed receptor modulation shapes network dynamics, with frontal regions necessary for global stabilization within the model but insufficient to explain the full pharmacological phenotype on their own. The findings point to receptor binding affinity as a key organizing principle linking molecular receptor organization to whole-brain function, though the authors note that the population-level receptor configurations derived from this single cohort will require validation across independent datasets, dosing regimens and target neuromodulatory systems before such in silico insights can achieve realistic clinical translation.

Subject of Research: Multi-receptor whole-brain modeling of atomoxetine's effects on resting-state brain dynamics

Article Title: Multi-receptor whole-brain modeling reveals affinity-driven chemoarchitectural mechanism of virtual atomoxetine action

Article References: Monge-Asensio, M., Demirtaş, M., Pfeffer, T., Zamora-López, G., Vohryzek, J., Sanz-Perl, Y., van den Brink, R. L., Donner, T. H., Kringelbach, M. L., & Deco, G. (2026). Multi-receptor whole-brain modeling reveals affinity-driven chemoarchitectural mechanism of virtual atomoxetine action. Frontiers in Computational Neuroscience, 20, 1854866. https://doi.org/10.3389/fncom.2026.1854866

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1854866

Keywords: atomoxetine, whole-brain modeling, catecholamines, receptor binding affinity, functional connectivity, resting-state fMRI, chemoarchitecture, computational neuroscience, norepinephrine, pharmacological fMRI, neuronal gain, brain dynamics